Abstarct: Medical science involves many challenges that require experts in this field to spend much time solving them. The behavior of neural neurons, one of the central cells forming the human brain and consisting of three main parts: the nucleus, dendrites, and axon, is one of the active research areas. The vital function of neurons in human brain activities, on the one hand, and their specific characteristics, such as not proliferating, on the other hand, make maintaining their health very necessary. One of the medical approaches to treating diseases related to these cells is using fluorescent microscopy images. The interconnectedness and communication of neuron components make it difficult for specialists to analyze these images. For this reason, approaches that can facilitate the analysis of the mentioned images can significantly improve the effectiveness of treatment methods. Segmenting the intracellular components of neurons in fluorescent microscopy images to improve and increase the speed of image analysis by specialists is one of the cases that can help treatment effectiveness. In this thesis, we seek to achieve this by using the deep learning approach and supervised learning paradigm with inaccurate labels, which previous researchers have addressed less. In previous methods, the presence of machine learning specialists alongside neuroscientists was required to determine the features and rules needed to segment cell components in neuron images. Utilizing new deep learning methods that have had favorable achievements in various scientific fields in recent years will minimize the cost of employing different specialists. The ultimate goal of neuron segmentation in images and their cellular components is to measure the length of each section. The results of this process will help analyze neuron behavior, such as shortening or lengthening a neuron or part of a neuron after applying a particular drug or treatment to it, and physicians and neuroscience specialists can easily access this information without spending too much time. For this purpose, a databaxse with inaccurate labels and a method that uses composite cost functions to train deep networks is used to compensate for the inaccuracy of the labels.